School of CSE Seminar Series- Yan Wang

Time: 
Friday, September 30, 2022 - 2:00pm to 3:00pm
Location: 
Atlanta, GA

Event Details

Speaker: Professor Yan Wang, George W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology
Date and Time: September 30, 2:00-3:00 p.m.
Location: Scheller, Room 102
Host: CSE Assistant Professor Victor Fung

Title: Physics-Informed Machine Learning for Physics-Based Data-Driven Design and Manufacturing

Abstract: When machine learning models are applied as the surrogate to solve complex engineering problems, the lack of training data due to the curse of dimensionality can make the surrogate predictions inaccurate. Physics-informed machine learning methods such as physics-constrained neural networks (PCNNs) have been introduced to integrate application-specific physical models in the data-driven surrogate to improve the training efficiency and reliability of predictions. Nevertheless, how to improve the convergence during the training of PCNNs with multiple loss terms in the total loss function is still challenging. In our recent work, an adaptive weight scheme was introduced to achieve better convergence. The training of PCNNs is also formulated as solving a minimax problem where the loss function for the worst-case scenario is minimized. A dual-dimer training algorithm is developed to solve the minimax problem. For uncertainty quantification, an efficient physics-constrained Bayesian neural network is also proposed to simultaneously reduce the bias and variance of predictions. The new PCNN framework has been applied in engineering design problems of heat transfer and phase transition. To improve the efficiency of data collection in physical experiments, we proposed a physics-constrained dictionary learning (PCDL) framework to solve the inverse problem of compressed sensing that is dedicated to manufacturing process monitoring. Data compression and the classification for diagnosis purpose can be done simultaneously in PCDL. 

Bio: Yan Wang, Ph.D. is a Professor of Mechanical Engineering and leads the Multiscale Systems Research Group at the Georgia Institute of Technology. The research areas of his group are in the intersection of design, manufacturing, and materials, and interested in multiscale modeling and simulation, uncertainty quantification, and physics-informed machine learning. He has co-authored over 200 peer-reviewed journal and conference papers, including the ones with best paper awards at the conferences of the American Society of Mechanical Engineers (ASME), The Minerals, Metals & Materials Society (TMS), the Institute of Industrial & Systems Engineers (IISE), and the International CAD Conference. He is a recipient of the U.S. National Science Foundation (NSF) CAREER Award, a National Aeronautics and Space Administration (NASA) Faculty Fellow, and an ASME Fellow. He currently serves as the Editor-in-Chief of the ASME Journal of Computing and Information Science in Engineering.